Blocks AI model Dessa aims to understand emotional signals
October 27, 2026 · 2 min read
Block, the company behind Square and Cash App, has introduced an AI language model named Dessa.
According to Block, this is the first AI model trained solely on human speech, entirely without text and built from the ground up. Dessa processes speech directly as audio, rather than first converting it into text. This allows it to recognize emotions, sarcasm, laughter, and pauses in real-time.
The model is full-duplex capable, meaning it can listen and speak simultaneously, and runs locally on a MacBook Pro M4 Max. It has been trained on just one single NVIDIA GB200 NVL72. Block cites three motives: curiosity, improved human-AI communication, and safety, as current language AIs ignore emotional signals. Dessa is not yet publicly available. The team is initially seeking AI safety researchers to test the model for unforeseen risks. Interested parties can apply through a form.
There have been previous attempts to create particularly emotionally sounding language AIs. The startup Hume AI introduced an AI voice with "emotional intelligence" named EVI in March 2024, followed by Kyutai's Moshi in September 2024. The Open Empathic initiative by LAION has been working since late 2023 to train empathic speech synthesis models using an open-source audio dataset.
OpenAI itself revised the "Advanced Voice Mode" of ChatGPT in June 2025 aiming to make the AI sound more emotionally nuanced and human-like, including improved intonation, pause placement, and the ability to express empathy and sarcasm more realistically.
It is important to note that sounding emotional is not synonymous with being emotional or intelligent, even if that impression may arise. For advanced language AIs, functions like reasoning during responses are more critical than superficial sound, as they truly impact answer quality.